|Publication number||US7606944 B2|
|Application number||US 11/746,994|
|Publication date||Oct 20, 2009|
|Filing date||May 10, 2007|
|Priority date||May 10, 2007|
|Also published as||US20080282030, WO2008140898A1|
|Publication number||11746994, 746994, US 7606944 B2, US 7606944B2, US-B2-7606944, US7606944 B2, US7606944B2|
|Inventors||George Alexander Kalwitz, Paul Andrew Ashmore|
|Original Assignee||Dot Hill Systems Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (16), Non-Patent Citations (1), Referenced by (10), Classifications (16), Legal Events (2)|
|External Links: USPTO, USPTO Assignment, Espacenet|
Methods and systems directed to automatically and dynamically determining sequential or random data access patterns for a storage controller are disclosed. In particular, methods and systems that optimize efficiency in accessing RAID storage devices according to sampled data access patterns are disclosed.
Redundant Array of Inexpensive Disk (RAID) storage systems are the predominant form of data storage in computer systems today that require high performance and/or high availability data storage for use in applications such as transaction processing, banking, medical applications, e-commerce, database applications, internet applications, mail servers, scientific computing, etc. A RAID system typically includes a number of physically independent disk drives coupled to a RAID controller. The RAID controller is a device that interfaces to a group of physical disk drives and presents them as a single logical disk drive (or multiple logical disk drives) to a computer operating system. RAID controllers employ the techniques of data striping and data redundancy to increase performance and data availability.
Data storage subsystems are commonly used to store data that can either be accessed randomly or in a sequential fashion, based on characteristics of the requesting application program that runs on a host computer. For example, transaction processing or OLTP programs tend to access data in a random fashion whereas a video server would tend to access data in a sequential fashion. Although hard disk drive storage devices can be used to store either random or sequential data, access efficiency is usually different for random or sequential workloads.
Hard disk drives are well known in the storage art, and have various latencies specific to their technology. There is rotational latency, which is the time it takes for the accessed disk platter to rotate such that the data to be accessed is beneath the magnetic head. There is also seek latency, which is the time it takes for the disk drive magnetic head to slew radially to a position where the data to be accessed is beneath the magnetic head. In addition, there are latencies associated with disk drive electronics and firmware to process incoming commands, manage the onboard disk drive cache memory, and send appropriate positioning commands to electromechanical mechanisms. The combination of the various latencies determines the data access time from incoming command to data processing completion (whether read or write).
Furthermore, hard disk drive devices have onboard buffers of varying size that cache read and write data. Storage controllers manage these buffers via queue depth and I/O size parameters. Maximum concurrent I/Os are the number of read or write commands that a disk drive can process simultaneously using onboard memory. It is technology and manufacturer dependent, and is a function of disk cache buffer size and disk drive design. It ranges from a minimum of one to a present-day maximum of 16 to 32 or higher. I/O size is usually highly variable, and can range from a single block being 512 bytes in size to a Megabyte or more. Storage controllers manage the number of concurrent I/Os to each storage device. Based on empirical testing, the number of allowed concurrent I/Os to an individual storage device is generally lower than the maximum concurrent I/Os supported. This number of allowed concurrent I/Os is called queue depth. Sequential workloads are generally optimized by utilizing a low queue depth and large I/O size, while random workloads are generally optimized by utilizing a higher queue depth and a small I/O size.
Some storage controllers are designed to operate in an entirely sequential or entirely random environment. They are set up to provide efficient access in a given mode all the time, without any concern about a changing workload that may alternate between random or sequential access. Such controllers may work well for their intended purpose, and are outside the scope of this invention. Many if not most storage controllers, on the other hand, are designed and intended to be used for general storage requirements—where the workload is unknown and may be a combination of sequential and random access. There are several approaches to dealing with an unknown workload.
A first approach is to optimize the controller for a sequential workload, where sequential accesses are handled efficiently and with minimal latency and random accesses are handled relatively inefficiently and with correspondingly higher latencies. This works well if all or most host accesses are sequential in nature and random accesses, if they occur, are allowed to be inefficient. However, if the workload changes to predominantly random host requests or if random accesses must also be conducted efficiently if they occur, then this scheme will not work well as it is only efficient for sequential accesses.
A second approach is to optimize the controller for a random workload, where random accesses are handled efficiently and with minimal latency and sequential accesses are handled relatively inefficiently and with corresponding higher latencies or lower bandwidth. This works well if all or most host accesses are random in nature and sequential accesses, if they occur, are allowed to be inefficient. However, if the workload changes to predominantly sequential host requests or if sequential accesses must also be conducted efficiently if they occur, then this scheme will not work well as it is only efficient for random accesses.
A third approach is to optimize a controller for neither sequential nor random workloads, but rather, for a compromise or average of sequential and random workloads. This has the advantage of favoring neither sequential nor random accesses, but averages inefficiency among both workloads. While this yields improved performance for mixed sequential and random workloads over the other two approaches, it does not handle either sequential or random workloads as efficiently as possible.
Accordingly, it would be advantageous to have a storage controller that automatically and dynamically optimizes operations to individual physical storage devices which form the RAID arrays according to actual sequential or random workloads.
In one aspect, the present invention provides a method of determining whether processed I/O requests represent sequential or random data access patterns. Incoming I/O requests are processed by RAID software, which organizes the data for efficient reads and writes to storage devices. The range of Logical Block Addresses (LBAs) of an I/O request processed by RAID software is compared to a predetermined setting for stripe size in order to determine the locality by chunk of the processed I/O request to the chunk location of the immediately previous I/O request to the same storage device. Comparing the chunk locations of these two I/O requests will determine whether the data access pattern for a particular storage device is sequential or random. This data access pattern may be sampled in order to create a storage device access profile. The sampling may be on every I/O request, once every nth I/O request, once every time interval, or based on the more complex operation of a state machine. A storage device access profile for each storage device is stored in memory, and is used to determine I/O size and queue depth parameters.
In another aspect, the present invention provides a storage controller for optimizing sequential or random accesses to storage devices based on address characteristics of processed I/O requests. First and second I/O requests are analyzed for locality of reference by a CPU or microprocessor. The chunk locations of the first and second I/O request are stored in a volatile or non-volatile memory within the storage controller. The locality of reference is sampled by the microprocessor to determine a storage device access profile for each storage device, which is also stored in the memory. The sampling is performed by the microprocessor according to regular event or time-based criteria. The storage device access profile is then used by the microprocessor to establish queue depth and I/O size parameters for I/O operations to storage devices through a device interface. The optimal queue depth and I/O size parameters are dependent on whether sequential or random accesses are occurring, and the characteristics of the specific storage devices used. By dynamically adjusting the queue depth and I/O size parameters to storage devices, I/O requests are processed efficiently whether the received workload corresponds to sequential or random accesses.
Other features and advantages of the present invention will become apparent upon study of the remaining portions of the specification and drawings.
Storage controllers store data on storage devices in a fashion that is optimized for rapid reading and writing by the storage controller. Each data I/O operation to storage devices is composed of a variable number of data blocks, where data block size is usually fixed at 512 bytes. However, it is possible that block size could be different than 512 bytes; 4K bytes, for example. Storage controllers organize the storage devices into contiguous equal-sized regions across all storage devices that make up a logical volume. These equal-sized regions are called Stripes, and are usually 64Kbytes. However, a Stripe could be as small as 16Kbytes or as large as 128Kbytes or 256Kbytes. The portion of a Stripe on a single storage device is called a Chunk.
Chunks are computed within the storage controller by utilizing information within SCSI CDBs (Command Descriptor Blocks). Specifically, the LUN identifies the individual disk drive (Host bus adapters and non-RAID applications) or group of disk drives (RAID controllers) being addressed. The LBA and block count provide addressing information to determine the specific stripe(s) and chunk(s) that the I/O request is being directed to. Relatively small I/O requests may fit within a single chunk and stripe on a single storage device, whereas relatively large I/O requests may be spread across multiple chunks and stripes on several disk drives.
It is this geometry of stripes and chunks that determines whether data accesses to a given storage device are sequential or random for a pair of accesses. Temporally consecutive I/O operations that address the same chunk on the same storage device are defined as sequential. Also, temporally consecutive accesses to chunks that are adjacent on the same disk drive are defined as sequential. All other temporally consecutive accesses to the same storage device are defined as random. RAID logical volumes utilize groups of disk drives, commonly 2 to 16, that are striped across all drives in the volume. Because RAID logical volumes contain multiple disk drives, it is necessary to constantly determine and track the sequential/random state of accesses to each drive in a RAID logical volume.
Although a simple approach to determining random or sequential behavior of I/O requests would be to scan incoming LBAs through the host I/O interface, this method does not always provide good accuracy. That is because host computer file systems frequently re-order host I/O requests such that temporally consecutive requests may appear random at the host interface while at a macro level the requests are actually sequential (or vice versa). Therefore, the determination of random or sequential behavior must be made after the storage controller software has put the storage controller I/O requests into a logically efficient order to maximize performance and minimize disk drive seek operations.
Once the sequential or random nature of I/O accesses are known, it is advantageous to adjust certain parameters involving communication with disk drives in order to maximize performance in either sequential or random environments. Specifically, the queue depth and I/O size written to disk drives can have a large bearing on reading or writing efficiency, which translates into performance. For sequential accesses, it is desirable to use a lower queue depth and large I/O size when accessing drives. For random accesses, it is desirable to use higher queue depth and small I/O size when accessing drives. Queue depth is a function of disk drive design, and can vary from a minimum of 1 to 16 or 32. I/O size is also variable, and commonly ranges from 512 bytes (i.e. one block) to a megabyte or more.
Alternatively, the storage controller 108 and storage device 112 may reside within the same storage subsystem 124, as shown in
Referring now to
In a simplest form, the locations may be sampled each time a new RAID I/O request is generated. Based on locality of the previous location to the current location, the sequential or random state for storage device access profile 504, 508, 512 would be updated. This would result in possibly rapid changes to the storage device access profiles, if the sampled locations of I/O requests were alternating between sequential and random. One improvement to this might be to sample every m I/O requests, where m was greater than 1 but small enough so that the storage controller would still respond rapidly to changing conditions. This approach has the benefit of filtering rapidly changing inputs such that the controller would tend to follow the inputs with a slower rate of change and possibly improved efficiency of storage device access. An alternative approach would be to sample at some regular time interval. This would also have the effect of filtering rapidly changing inputs such that the controller would tend to follow the inputs with a slower rate of change.
Considering the three variables pertaining to a single storage device n, although the storage device access profile 504 may be sequential or random at the same time as the chunk locations 516, 520 associated with that same storage device, it may not be. That is because sampling is applied to the chunk locations such that the sampling behavior will determine when the storage device access profile 504 is updated. For simple cases of sampling, the method applied is shown in
Sequential or random accesses are determined separately for each storage device. Referring now to
Referring now to
Referring now to
The invention includes a process for determining the current storage device access profile for each storage device 112, and this is shown in
In the simplest case where the chunk locations to the last two temporally consecutive RAID I/O requests are sampled whenever a new RAID I/O request is issued to a storage device, this determination would always be affirmed and control would proceed to step 724. However, there are two other cases in which the sample criteria might not initially be met, with the result that control would return to step 708 to await a new RAID I/O request. The first such case would be if the sampling criteria involved sampling every m RAID I/O requests to a given storage device. Each time step 720 was encountered, m would be decremented. Upon reaching zero, the count would be restored to m and flow would proceed to step 724. The second such case would be if the sampling criteria involved sampling at a regular time interval. This would require a separate timer to be operated. Similar to the method described in the first such case, each time step 720 were encountered, there would be a check to verify if the timer had reached zero. Upon reaching zero, the timer would be restored to begin a new count and flow would proceed to step 724.
At step 724, a determination is made as to whether the sampled chunk locations represent either a sequential or random condition. If sequential, then flow will proceed to step 728, where the storage device access profile is updated to sequential, and flow returns to step 708 to wait for a new I/O request. If random, then flow will proceed to step 732, where the storage device access profile is updated to random, and flow returns to step 708.
In the preferred embodiment, a predetermined number of samples having the opposite state as the current storage device access profile would be required before the storage device access profile changed to the opposite state. This is reflected in
It should be noted that it would be just as valid for the state machine to power up into random state −2 820 as sequential state 2 804. In addition, the sample frequency may be based on either individual RAID I/O requests or time. That is, while the preferred embodiment would be based on sampling every RAID I/O request to the given storage device, the state machine could be mechanized to sample every nth RAID I/O request, or at some regular time interval. While the alternative sampling approaches would respond more slowly than the preferred embodiment, they would have the benefit of less processor 216 utilization and slower transitions in the event of statistically noisy inputs characterized by numerous and rapid transitions between random and sequential RAID I/O requests. The state machine of
In conjunction with
While the discussion to this point has focused on what would constitute sequential or random RAID I/O requests, and when transition to a different storage device access profile occurs, there now must be definition of the characteristics of the storage device access profile. As stated earlier, sequential access profiles have correspondingly low queue depth and large I/O sizes relative to random access profiles. Although all disk drives have a minimum queue depth of 1, it is advantageous to use a queue depth of 2 for all sequential profiles (except SATA I drives), since there is command processing latency benefit to having the next I/O ready to be executed by the drive at all times. Since SATA I drives only support a queue depth of 1, that would have to be used for all profiles. The sequential access profile I/O size is heavily drive-dependent, and depends on the design of the individual disk drive and memory available on the drive. Empirical testing may be needed to determine the optimal I/O size to use, although efficient values in the area of 1 MByte for Fibre Channel and SATA II drives, 64 Kbytes for SATA I drives, and 512 Kbytes for SAS drives may be used.
In comparison to sequential storage device access profiles, random storage device access profiles would generally be characterized by higher queue depths. However, even though some drives support higher queue depths of 32 or even 64 commands, those figures may not yield the best performance. Again, empirical testing would be necessary to confirm optimal values for any particular disk drive. An efficient queue depth of 8-12 commands may be used for some disk drives. Storage device command I/O sizes are most efficiently dealt with when they match the size of the RAID I/O request. For example, for a 4 Kbyte RAID I/O request, the I/O size to storage devices should be 4 Kbytes.
The foregoing discussion of the invention has been presented for purposes of illustration and description. Further, the description is not intended to limit the invention to the form disclosed herein. Consequently, variations and modifications commensurate with the above teachings, within the skill or knowledge of the relevant art, are within the scope of the present invention. The embodiments described herein above are further intended to explain the best mode presently known of practicing the invention and to enable others skilled in the art to utilize the invention in such or in other embodiments and with the various modifications required by their particular application or use of the invention. It is intended that the appended claims be construed to include alternative embodiments to the extent permitted by the prior art.
Finally, those skilled in the art should appreciate that they can readily use the disclosed conception and specific embodiments as a basis for designing or modifying other structures for carrying out the same purposes of the present invention without departing from the spirit and scope of the invention as defined by the appended claims.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US4870401 *||Mar 2, 1988||Sep 26, 1989||Dallas Semiconductor Corporation||Electronic key locking circuitry|
|US5426736||May 26, 1993||Jun 20, 1995||Digital Equipment Corporation||Method and apparatus for processing input/output commands in a storage system having a command queue|
|US5522054||Sep 13, 1993||May 28, 1996||Compaq Computer Corporation||Dynamic control of outstanding hard disk read requests for sequential and random operations|
|US5613155||Jun 7, 1995||Mar 18, 1997||International Business Machines Corporation||Bundling client write requests in a server|
|US5625794 *||Feb 24, 1993||Apr 29, 1997||Hitachi, Ltd.||Cache mode selection method for dynamically selecting a cache mode|
|US6167496||Feb 18, 1998||Dec 26, 2000||Storage Technology Corporation||Data stream optimization system for video on demand|
|US6170042 *||Mar 19, 1998||Jan 2, 2001||Seagate Technology Llc||Disc drive data storage system and method for dynamically scheduling queued commands|
|US6427184||Jun 1, 1998||Jul 30, 2002||Nec Corporation||Disk drive with prefetch and writeback algorithm for sequential and nearly sequential input/output streams|
|US6687765 *||Jan 16, 2001||Feb 3, 2004||International Business Machines Corporation||System, method, and computer program for explicitly tunable I/O device controller|
|US6879939||Dec 27, 2002||Apr 12, 2005||Unisys Corporation||Control of distribution of input/output operations during system testing|
|US6922414||Aug 21, 2000||Jul 26, 2005||Hewlett-Packard Development Company, L.P.||Apparatus and method for dynamic command queue depth adjustment for storage area network nodes|
|US7089347||Aug 5, 2003||Aug 8, 2006||Hitachi, Ltd.||Computer system for managing performances of storage apparatus and performance management method of the computer system|
|US20030088734 *||Nov 2, 2001||May 8, 2003||Cavallo Joseph S.||Transaction request coalescing process|
|US20050108450 *||Mar 18, 2004||May 19, 2005||Hirofumi Sahara||Information processing system and method|
|US20060064572 *||Sep 8, 2005||Mar 23, 2006||Adaptec, Inc.||Automatic initialization|
|US20070055798||Aug 31, 2005||Mar 8, 2007||Ain Jonathan W||Apparatus and method to adjust one or more input/output parameters for a computing system|
|1||*||Smith, Alan Jay "Line (Block) Size Choice for CPU Cache Memories", Jun. 1985, UC Berkeley.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US7836107 *||Dec 20, 2007||Nov 16, 2010||Microsoft Corporation||Disk seek optimized file system|
|US8239589 *||Mar 31, 2010||Aug 7, 2012||Amazon Technologies, Inc.||Balancing latency and throughput for shared resources|
|US8473646||Jun 21, 2012||Jun 25, 2013||Amazon Technologies, Inc.||Balancing latency and throughput for shared resources|
|US8713219 *||Jan 26, 2011||Apr 29, 2014||Fujitsu Limited||Switching device, switch control method, and storage system|
|US8972689 *||Feb 2, 2011||Mar 3, 2015||Violin Memory, Inc.||Apparatus, method and system for using real-time performance feedback for modeling and improving access to solid state media|
|US9396067 *||Apr 18, 2012||Jul 19, 2016||American Megatrends, Inc.||I/O accelerator for striped disk arrays using parity|
|US20090164535 *||Dec 20, 2007||Jun 25, 2009||Microsoft Corporation||Disk seek optimized file system|
|US20110191519 *||Jan 26, 2011||Aug 4, 2011||Fujitsu Limited||Switching device, switch control method, and storage system|
|US20110246677 *||Apr 5, 2010||Oct 6, 2011||Lsi Logic Corporation||Systems and methods for controlling commands for target devices|
|US20150242133 *||Feb 21, 2014||Aug 27, 2015||Lsi Corporation||Storage workload hinting|
|U.S. Classification||710/5, 710/52, 711/114, 711/100, 710/15, 711/158|
|International Classification||G06F13/00, G06F7/00|
|Cooperative Classification||G06F3/0613, G06F3/0689, G06F3/0656, G06F3/0659|
|European Classification||G06F3/06A4T6, G06F3/06A4T2, G06F3/06A2P4, G06F3/06A6L4R|
|May 25, 2007||AS||Assignment|
Owner name: DOT HILL SYSTEMS CORPORATION, CALIFORNIA
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:KALWITZ, GEORGE ALEXANDER;ASHMORE, PAUL ANDREW;REEL/FRAME:019345/0557
Effective date: 20070510
|Apr 19, 2013||FPAY||Fee payment|
Year of fee payment: 4